Application Of Neural Network To Technical Analysis Of Stock Market Prediction

FROM - MON APR 27 09:57:18 1998 Page 1 of 14 

Application Of Neural Network To Technical 

Analysis Of Stock Market Prediction 

Hirotaka Mizuno, Michitaka Kosaka, Hiroshi Yajima 

Systems Development Laboratory 

Hitachi, Ltd. 

8-3-45 Nankouhigashi, Suminoe-Ku 

Osaka 559-8515 

JAPAN 

Norihisa Komoda 

Department of Information Systems Engineering 

Faculty of Engineering, Osaka University 

2-1 Yamadaoka, Suita 

Osaka 565-0871 

JAPAN 

Abstract: This paper presents a neural network model for technical analysis of stock market, and its 

application to a buying and selling timing prediction system for stock index. When the numbers of 

learning samples are uneven among categories, the neural network with normal learning has the 

problem that it tries to improve only the prediction accuracy of most dominant category. In this paper, 

a learning method is proposed for improving prediction accuracy of other categories, controlling the 

numbers of learning samples by using information about the importance of each category. 

Experimental simulation using actual price data is carried out to demonstrate the usefulness of the 

method. 

Keywords: stock market prediction, technical analysis, neural network, learning method 

Hirotaka Mizuno received his BE and ME degrees from Osaka University in 1979 and 1981, 

respectively. He is currently Senior Research Engineer at Systems Development Laboratory, Hitachi, 

Ltd., Osaka. His research interests include pattern processing, and decision support systems. He is a 

member of the Information Processing Society of Japan, and of the Institute of Electrical Engineers of 

http://www.ici.ro/ici/revista/sic98_2/art03.html 7/12/2001


Japan. 

Michitaka Kosaka received his BE, ME, and Ph.D degrees from Kyoto University in 1975, 1977 and 

1984, respectively. He is currently Department Manager of the 1st Research Department at Systems 

Development Laboratory, Hitachi, Ltd., Osaka. His main research interest is in large scale system 

engineering. He is a member of the Institute of Electrical Engineers of Japan, and of the Society of 

Instrument and Control Engineering. 

Hiroshi Yajima received his BE, ME, and Ph.D degrees from Kyoto University in 1973, 1975 and 

1992, respectively. He is currently Department Manager of Kansai Systems Laboratory at Systems 

Development Laboratory, Hitachi, Ltd., Osaka. His research interests include information systems 

planning and decision support systems. He is a member of the IEEE, the Information Processing 

Society of Japan, and of the Institute of Electrical Engineers of Japan. 

Norihisa Komoda received his BE, ME, and Ph.D degrees from Osaka University in 1972, 1974, and 

1982, respectively. He is currently Professor at the Department of Information Systems Engineering, 

Faculty of Engineering, Osaka University. He stayed at the University of California, Los Angeles, as 

a visiting researcher from 1981 to 1982. His research interests include systems engineering and 

knowledge information processing. He is a member of the IEEE, the ACM, the Institute of Electrical 

Engineers of Japan, and of the Society of Instrument and Control Engineering. He received the 

Awards for outstanding paper and the Awards for outstanding technology both from the Society of 

Instrument and Control Engineering. 

1. Introduction 

This paper proposes a neural network model for technical analysis of stock market, 

and its learning method for improving the prediction capability. In stock market 

prediction, many methods for technical analysis have been developed and are being 

used [1]. In technical analysis, technical indexes calculated from price sequence are 

used to predict the trend of future price changes. Many statistical methods have been 

proposed, but the results are insufficient in prediction accuracy. In this paper, neural 

network [2, 3] is applied to technical analysis as a prediction model, and a buying and 

selling timing prediction system for TOPIX (Tokyo Stock Exchange Prices Index) is 

presented. TOPIX is a weighted average of prices of all stocks listed on the First 

Section of Tokyo Stock Exchange. 

Several neural network models have already been developed for market prediction. 

Some are applied to predicting future price or rate of changes [4], and some are 

applied to recognizing certain price patterns that are characteristic of future price 

changes [5]. In these models, however, little is considered about the learning method 

of neural network. In case the numbers of learning samples are uneven among 



categories, neural network models with normal learning try to improve only the 

prediction accuracy of the most dominant category which might be less important than 

others. 

This paper proposes a learning method that contributes to improving prediction 

accuracy of the other categories, which are more important. In the method, the 

numbers of learning samples are controlled by using information about the importance 

of each category. Experimental simulation using actual TOPIX price data is carried 

out to demonstrate the usefulness of the proposed method. 

2. TOPIX Prediction System 

The overview of the proposed buying and selling timing prediction system for TOPIX 

is shown in Figure 1. The prediction system classifies the input pattern that consists of 

several technical indexes of TOPIX, and generates a buying or selling timing signal 

for notifying users [6]. As shown in the Figure, the system consists of a neural 

network, a preprocessing unit, and a postprocessing unit. The preprocessing unit 

normalizes each technical index into an analog value in 0 to 1 to form an input pattern 

into the neural network. Then the network recognizes the turning point of the TOPIX 

price curve from the input pattern. Finally, the postprocessing unit converts the result 

of recognition into a buying and selling timing signal. 

3. Neural Network Model 

3.1 Network Architecture 

Figure 1. Overview of TOPIX Prediction System 



As shown in Figure 1, the neural network as prediction model is a hierarchical 

network that consists of three layers: the input layer, the hidden layer, and the output 

layer. Each unit in the network is connected to all units in the adjacent layers. Each 

unit receives outputs of the units in the lower layer and calculates the weighted sum to 

determine total input. Then the output is determined by applying the logistic function 

[7] to the total input. As a result, the output ranges in 0 to 1. 

3.2 Input Data Selection 

Data items to form the input pattern to the system are technical indexes of TOPIX. 

Typical technical indexes are: 

Moving averag 

This is an average of the prices over certain past period, and developed for 

allowing users to understand the trend without everyday fluctuation. There are 

several variations according to the period: 6, 10, 25, 75, 100, 150, and 200 days. 

The direction and position of moving average curve are used for predicting the 

price change. 

Deviation of price from moving average 

This index is used for checking whether the price on each day is too high or too 

low if compared with the expected price. 

Psychological line 

This index is calculated by dividing the number of days of price ups by certain 

past period. This is used for predicting the price change from the rhythm of ups 

and downs. 

Relative strength index 

This index is similar to the psychological line, and is calculated by dividing the 

sum of price ups by the sum of price ups and downs over certain past period. 

Each of these indexes is normalized into 0 to 1 to form an input pattern to the neural 

network model. 

3.3 Output Data Definition 

In the neural network model, the output layer has three units. As output patterns of the 

network, we define three patterns as shown in Table 1. Each corresponds to specific 

TOPIX curve patterns: buying signal (i.e. current price is at bottom), selling signal 

(i.e. current price is at top), and no-change (i.e. otherwise), respectively. Bottoms and 

tops in the price curve are closely related to buying and selling timings. When 

teaching the neural network model, each correct output pattern of learning sample is 

calculated from three TOPIX data as described in Table 1: current price, price at five 



weeks before, and price at five weeks later. 

Table1 : Relation between TOPIX Curve and Correct Output Pattern of 

Neural Network 

Figure 2 shows an example of computer generated correct buying and selling signals 

on the TOPIX graph. Buying and selling signals are designated by black circles and 

white triangles, respectively. These correct signals are calculated from three TOPIX 

data (i.e. current price, price at five weeks before, and price at five weeks later) as 

described in Table 1, and are not recognized by human experts. However, an expert 

analyst comments that these correct signals are almost satisfactory as long as the 

investment period is supposed to be of three months. 

Figure 2. Example of Computer Generated Correct Outputs 

Because the output of the units in the output layer ranges in analog of 0 to 1 in a 

neural network, an actual output pattern may not match with any of the three patterns. 

In this case, the postprocessing unit converts the analog value to 0 or 1 by using two 

thresholds. In the experimental simulation further described , 0.4 and 0.6 are used as 

thresholds. When the output is beneath 0.4, then 0 is selected. In the same way, when 

the output is more than 0.6, then 1 is selected. If the output is between 0.4 and 0.6, or 

if the converted pattern still does not match any of the three categories, the system 

notifies that prediction has failed. 



4. 

Figure 3. Equalized Learning Procedure 

Equalized Learning Method 

In the prediction system, a learning method is developed for improving prediction 

accuracy. As earlier described , the neural network model tries to generate one of 

three output patterns for classification. 

Learning of the network is done using the back- propagation algorithm [7]. The 

learning strategy in this algorithm is to modify the weights of connections between 

units towards decreasing the total sum of prediction errors. However, in case that the 

deviation of the numbers of learning samples is large among output pattern categories, 

the following problem arises: the neural network will have the tendency of trying to 

improve only the prediction accuracy of the most dominant category. As a result, the 

prediction system generates less buying and selling timing signals than expected. This 

is because little is considered about the difference of importance among categories. 

We must teach the neural network that we expect it to generate buying and selling 

timing signals, and that generating no buying signal or no selling signal makes no 

sense. 

To overcome this problem, we propose the "equalized learning method" [8] as shown 

in Figure 3. Before learning samples are fed to the neural network, histogram 

calculation step and sample duplication step are inserted. Equalized learning is carried 



out in the following way: 

1. Histogram calculation 

Histogram is calculated by counting the learning samples in each category. 

2. Sample duplication 

By using information about the importance of categories and the histogram, 

learning samples in each category are duplicated to modify the histogram 

3. Shuffling new samples 

4. Feeding to the neural network 

In the sample duplication step, the more important the category is, the larger the 

coefficient of duplication assigned. In the TOPIX prediction system, learning samples 

of buying signal and selling signal are less in spite of the fact that these two categories 

are more important than the no-change category. Therefore, the coefficients of 

duplication are set to higher values of these categories. However, it is difficult to 

decide these values for each category, though it is possible to decide their order. In the 

prediction system, the coefficients are set so that the histogram becomes nearly flat. 

Consequently the prediction system begins to generate more buying and selling 

signals at proper timings. 

5. Experimental Simulation 

5.1 Data and Method 

To verify the accuracy of the prediction system, we have carried out an experimental 

simulation applied to actual weekly TOPIX data. Figure 4 shows the test data. As 

shown in the Figure, 260 weekly data from September 1982 to August 1987 (five 

years) have been used as learning samples for making prediction models. Following 

119 weekly data from October 1987 to January 1990 (two years) are used as 

prediction samples for evaluation. For both samples, correct output patterns are 

calculated according to the definition described in Table 1. 

Eleven technical indexes of TOPIX are selected to form input patterns into the system 

as listed in Table 2. These indexes are selected by a human expert analyst. Each of 

these has been commonly used and believed useful in technical analysis. 

Table 2. Technical Indexes To Form Input Pattern 



Table 3 shows the details of the computer generated correct output patterns of the test 

data. As for the learning samples, ratios of the samples in three categories (i.e. selling 

signal, buying signal, and no-change) are 16 %, 15 %, and 69 %, respectively. As for 

prediction samples, ratios of samples are 26 %, 24 %, and 50 %, respectively. In both 

samples, the samples in no-change category are extremely larger than those in the 

other two categories. This suggests that a neural network with normal learning method 

may always generate no-change category as its output pattern. 

Table 3. Details of Correct Output Patterns in theTest Data 

Figure 4 also shows that both the learning period and the simulation period have an 

increasing trend on the whole. This suggests possible difficulty in predicting selling 

timings because of the lack of efficient learning samples for selling signals. 

Figure 4. Test Data for Experimental Simulation (TOPIX Weekly Data) 

For comparison of prediction performance, three prediction models are applied: 



1. neural network model with the proposed equalized learning 

coefficients of duplicating learning samples: 10, 2, and 11 

11 units in input layer, 10 units in hidden layer, and 3 units in output layer 

coefficients for updating weights: 0.1 and 0.9 

5,000 iterations 

thresholds for postprocessing: 0.4 and 0.6 

2. neural network model with normal learning 

same parameters as the above model except no duplication of learning samples 

20,000 iterations 

3. statistical model based on discriminant analysis [9, 10] 

two models are used: one for classifying whether buying signal or not, and one 

for classifying whether selling signal or not 

Mahalanobis' generalized distance is used as distance measure 

5.2 Result of Prediction 

Tables 4, 5, and 6 show the results of learning and prediction by the three prediction 

models. Table 7 summarizes these results. 

Table 4. Prediction Result by Neural Network Model with Equalized Learning 

Table 5. Prediction Result by Neural Network Model with Normal Learning 



Table 6. Prediction Result by Statistical Model 

Table 7. Performance Comparison of Prediction 

As shown in Table 5, the neural network model with normal learning has learned the 

learning samples in no-change category completely. Total ratio of correctness for 

learning is 97 % and is sufficient. However, as for learning selling signal samples, the 

ratio of correctness is 83 % (35/42), lower than the total one. That means that the 

network learns mainly the samples in no-change category, which is most dominant of 

the three categories. As for prediction, though the ratio of correctness for predicting 

no-change is 83 % (49/59), the ratios for predicting buying signals and selling signals 

are 66 % (19/29) and 16 % (5/31), i.e. much lower. It can be understood that not 

enough learning has been carried out for the learning samples in these two latter 

categories. 



On the other hand, as shown in Table 4, though the neural network model with the 

equalized learning method learned the learning samples in no-change category for 88 

% (157/178), it learned the selling signals and the buying signals in 98 % (41/42) and 

100 % (40/40). These exceed the performance of neural network model with normal 

learning by 15 % and 2 %, respectively. This indicates that learning was carried out 

equally over the three categories. As for prediction, though the ratio of correctness for 

predicting no-change is 66 % (39/59) and lower than the normal learning model, the 

ratio of correctness for predicting selling signals is 45 % (14/31), which exceeds the 

normal learning model by 29 %. The ratio of correctness for predicting buying signals 

is 76 % (22/29) and exceeds the normal learning model by 10 %. The total ratio of 

correctness is 63 % and there is no significant difference if compared with the normal 

model. Assuming that the improvement of prediction accuracy for selling and buying 

signals is more important to users, these results demonstrate that the proposed neural 

network model shows higher performance of prediction than the normal neural 

network model does. However, since the prediction accuracy for selling signals is still 

low, further analysis is necessary to improve this. 

As shown in Table 6, the performance of the statistical model is much lower than that 

of the normal neural network model in every aspect except for the prediction accuracy 

for selling signals. 

In Figure 5 (a), the buying and selling timing signals generated by the prediction 

system are plotted. Comparison with the computer generated correct signals in Figure 

5 (b) shows that the signals by the system are similar to the correct ones on the whole. 

An expert analysis comments that the signals by the system are almost satisfactory 

except that proper signals do not appear in some situations. This means the selling 

timing signals by the system are still less satisfactory than expected. 

Figure 5. Comparison of Buying and Selling Signals 

5.3 Result of Buying and Selling Simulation 

To verify the usefulness of the proposed prediction system, buying and selling were 

simulated. For comparison of performance reason, two other cases of simulation have 



also been carried out using a single technical index: psychological line, and relative 

strength index (RSI) as listed in Table 8. These indexes have been used everyday in 

stock market analysis. Buying and selling strategy in each case is also shown in the 

Table. One more case of buy-and-hold strategy was also taken as a basis. In buy-andhold, 

buying is done once at the beginning of the simulation and selling is done at the 

end of it, that means no active buying and selling. Prediction systems are required to 

achieve higher performance than this strategy's. 

Table 8. Simulation Models and Buying-and-Selling Strategies 

As a measure of performance, rate of profit is used, calculated as follows: 

at buying situations: 

at selling situations: 

The simulation results are summarized in Table 9. From these results, an expert 

analyst will make the following comments: 

Table 9. Performance Comparison of Buying-and-Selling Simulation 



Buying situations: 

Before converting in yearly profit rate, the proposed neural network model 

scored highest performance (1.51). On the other hand, in yearly profit rate, 

psychological line and RSI showed better performances. That means the neural 

network model does not generate proper buying and selling signals at the best 

timings. However, since it exceeds the performance achieved by the buy-andhold 

strategy, the timings of signals by the proposed neural network model are 

proved to be useful on the whole. 

Selling situations: 

Rate of profit must be higher than one if the selling timing signals are 

appropriate. How-ever, the performance achieved by the proposed neural 

network model is 0.99 and slightly under one. That means the model does not 

generate appropriate selling signals. However, if comparing the result with other 

cases (0.84 and 0.90), the neural network model minimized the loss. In this 

sense, the proposed model is relatively the best of the three cases. 

Overall: 

On the whole, although the neural network model did not make the higher profit 

in buying situations than in other cases, it made the highest in selling situations 

in the sense of a minimum loss. Consequently, its overall performance (1.20) was 

greater than that achieved by single use of the technical index (0.94, 1.05). 

However, it was still slightly lower than that of buy-and-hold strategy (1.21). 

This might be because the trend of price change is increasing throughout the 

learning and prediction period in this simulation. 

6. Conclusion 

This paper proposed a neural network model for the technical analysis of stock market 

prediction, and its application to the buying and selling timing prediction system for 

TOPIX. So far, little has been considered about the learning method of neural 

network. In case the numbers of learning samples are uneven among categories, neural 

network models with normal learning try to improve only the prediction accuracy of 

the most dominant category that may be less important than others. This paper 

proposed a learning method that contributed to improving prediction accuracy of 

other, more important categories. In the method, the numbers of learning samples are 

controlled by using information about the importance of each category. Experimental 

simulation applied to practical data has demonstrated that the prediction system 

generates buying and selling signals at more proper timings on the whole, and made 

higher profit compared with that yielded by a single use of each technical index. As 

future work, we have to study several problems which will help improve the model. 

First, we have to study what makes it difficult for the system to generate selling 



signals at appropriate timings. Second, on analyzing the network, it is necessary to 

know the effectiveness of each technical index used as input. 

REFERENCES 

1. PRING, M.J., Technical Analysis Explained, MCGRAW-HILL, 1985. 

2. BENARD, W., DAVID, E.R. and MICHAEL, A.L., Neural Networks: 

Applications in Industry, Business and Science, COMMUNICATIONS OF 

THE ACM, Vol. 37, No. 3, 1994, pp. 313-333. 

3. WASSERMAN, P.D., Neural Computing: Theory and Practice, VAN 

NOSTRAND REINHOLD, 1989. 

4. KIMOTO, T., ASAKAWA, K., YODA, M. and TAKEOKA, M., Stock Market 

Prediction System with Modular Neural Networks, Proceedings of the 1990 

International Joint Conference on Neural Networks, 1990, pp. 1/1-1/6. 

5. KAMOJO, K. and TANIGAWA, T., Stock Price Pattern Recognition - A 

Recurrent Neural Network Approach, Proceedings of the 1990 International 

Joint Conference on Neural Networks, 1990, pp. 1/215-1/221. 

6. KOSAKA, M., MIZUNO, H., SASAKI, T., SOMEYA, R. and HAMADA, N., 

Applications of Fuzzy Logic/Neural Network to Securities Trading Decision 

Support System, Proceedings of the 1991 IEEE International Conference on 

Systems, Man, and Cybernetics (SMC '91), 1991, pp. 1913-1918. 

7. RUMELHART, D.E., HINTON, G.E. and WILLIAMS, R.J., Learning Internal 

Representation by Error Propagation, in D.E. Rumelhart et al [Eds.] Parallel 

Distributed Processing , MIT PRESS, 1986, pp. 1/318-1/362. 

8. MIZUNO, H., KOSAKA, M. and YAJIMA, H., Stock Market Prediction 

System Based On Technical Analysis of Stock Index Using Neural Network, 

Proceedings of the 15th IMACS World Congress on Scientific Computation, 

Modelling and Applied Mathematics (IMACS '97), 1997, pp. 4/49-4/54. 

9. BOLCH, B.W. and HUANG, C.J., Multivariate Statistical Methods for 

Business and Economics, PRENTICE-HALL, 1974. 

10. LACHENBRUCH, P.A., Discriminant Analysis, HAFNER, 1975.

Application Of Neural Network To Technical Analysis Of Stock Market Prediction

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?